The college football early season UNDER. Frozen before kickoff, graded in public, and it failed.
One college football totals angle cleared our historical gate, and it could only go live if the 2026 forward test held. This is that angle, published with its exact rule before the test started. Across 2026 weeks 1-3 we graded every qualifying game against the closing number. The final count is 62 under, 69 over, 2 push: 47.3% under on n=131, below the 52.38% gate. We said pass or fail, the result stands in public. It failed, and here it stands.
the weeks 1-3 window closed Sep 20, 2026 · the angle is off the board
The final result · 2026 forward test · closed Sep 20, 2026
62-69-2 (under-over-push), n=131 · 47.3% under · gate 52.38% · failed
133 games qualified at a closing total of 47 or higher; 2 pushed, leaving n=131. 8 of 157 completed FBS games had no pre-kickoff closing line in our capture and were skipped, not graded. One correction is on the record: until Sep 17 the grader read a lines table that lost pre-kickoff captures, so the early count covered only 34 games. It was repointed at the full capture history that day, the sample moved to 85, and every week in the final count is graded from the full history.
What it means. 47.3% on n=131 is below the 52.38% gate, so the angle failed the one test it could not see coming. At this sample the result is not distinguishable from a coin flip, and a coin flip loses money at standard pricing. One three week window cannot prove the historical 54.75% (n=884) was never real, and it does not have to: the rule was that the angle had to clear the gate on games it had never seen, and it did not. The 47 point filter was chosen after looking at the data, and 2025 had already slipped to 52.1% (n=119). The plainest reading is that the backtest overstated the edge, or the market has since closed it. Either way the angle is off the board. It is no longer a signal on the game board, it is logged as rejected in The Lab's testing log, and the article that introduced it now carries the same final count.
The rule · exactly as registered, nothing tunable left
FBS regular season, weeks 1 through 3. Closing total of 47 or higher. The side is the UNDER, measured against the closing no-vig total. That is the whole rule.
Why it might be real: September totals get shaded upward by peak-season public over-action while preseason projections still carry a real information gap about which teams are actually good, and the market cannot fully correct before game time. High totals are where that shading concentrates. The mechanism's fingerprint is an early-vs-late season gap in under rates, and we track that fingerprint explicitly.
The history · 2018-2025, real closing lines, sample on every number
The cyan line is 52.4%, the rate a bettor needs at standard vig just to break even. A backtest clearing it on average is a reason to run a forward test, not a reason to bet.
What our own adversarial review flagged · read this before the history
The 47-point filter was found in the data, not predicted in advance. The registered claim covered all weeks 1-3 games, which came in at 53.4% and failed our 54% gate; the ≥47 cut is what cleared it, and it was selected after looking. That is exactly the kind of choice that overstates significance, which is why the rule is now frozen and 2026 is its first clean, no-touch test.
The most recent season is the weakest evidence. 2025 came in at 52.1% (n=119), below the breakeven gate, and the early-vs-late gap that carries the mechanism collapsed to half a point after running 8 to 10 points in 2023-2024. Either the market adapted, or 2025 was noise at n=119. We do not know, and neither does anyone selling certainty.
Season-to-season swing is wide: 50.9% to 59.6%. The average is driven by the strong years. That is why the gate has two stages: clearing the historical bar earned this angle a forward test, and it could go live only if the 2026 forward window graded out. It did not: 47.3% under on n=131.
What we said we would track · and what happened
- Every qualifying game (FBS, weeks 1-3, closing total ≥ 47), expected roughly 100-120 games, graded UNDER or not against the closing no-vig total. The running count posts with its n, misses included. What happened: 133 games qualified, 2 pushed, n=131 graded.
- The pass bar, stated now: at that sample the confidence floor only clears 52.4% breakeven if the under rate lands around 55% or better. Below that, the angle does not graduate, whatever the average says. What happened: 47.3% under on n=131. That misses the 55% bar and the 52.38% breakeven gate itself, so there is no reading of the bar under which it passed.
- The mechanism check: the early-vs-late gap must reopen (roughly 4+ points) for the causal story to survive. A pass on rate with no gap is a flag, not a confirmation. What happened: moot. The rate failed on its own, so the gap check decides nothing.
- 2020 stays excluded from the deployment case (n=33, anomalous COVID scheduling).
Grading ran from the first Week 1 final on Aug 29, 2026 through the last week 3 game on Sep 20, 2026. The list above is the text we published before the season, left as written.
Into the testing log at The Lab, with every other angle that came off the board. The last angle that reached this stage hit 56.7% across 598 backtest games and came off when its out-of-sample season reversed to 41.4% (n=70). This one got the same treatment: 54.75% across 884 backtest games, 47.3% on n=131 when it counted.
Get the research as it publishes
New research and graded results by email, the misses included. This page is what a miss looks like here.
The board, the numbers, and what we are playing. No spam. 21+. Unsubscribe anytime.
Provenance: hypothesis 64 (registered Jun 26, 2026 · validated Jul 1, 2026 on the ≥47 variant), experiments 55 and 58, adversarial review of Jun 30, 2026, all rows in the Edge Labs research database; verified against the live rows on Jul 2, 2026. Season numbers are experiment 58's primary cut. The final 2026 forward test count is experiment 112, graded by our public grader and re-run once on Sep 21, 2026 to confirm; hypothesis 64 is now rejected. Research and context, never a pick, 21+.